1550 Background: Dense breast tissue is an independent risk factor for malignancy and can mask cancers on mammography. Yet, radiologist-assessed mammographic breast density is subjective and varies widely between and within radiologists. Our deep learning (DL) model was implemented into routine clinical practice at an academic breast imaging center and was externally validated at a separate community practice, with both sites demonstrating high clinical acceptance of the model’s density predictions. The aim of this study is to demonstrate the influence our DL model has on prospective radiologist density assessments in routine clinical practice. Methods: This IRB-approved, HIPAA-compliant retrospective study identified consecutive screening mammograms without exclusion performed across three clinical sites, over two time periods: pre-DL model implementation (January 1, 2017 through September 30, 2017) and post-DL model implementation (January 1, 2019 through September 30, 2019). Clinical sites were as follows: Site A (the academic practice where the DL model was developed and was implemented in late 2017); Site B (an affiliated community practice which implemented the DL model in late 2017 and was used for external validation); and Site C (an affiliated community practice which was never exposed to the DL model). Patient demographics and radiologist-assessed mammographic breast densities were compared over time and across sites. Patient characteristics were evaluated using Wilcoxon test and Pearson’s chi-squared test. Multivariable logistic regression models evaluated the odds of a dense breast classification as a function of time period (pre-DL vs post-DL), race (White vs non-White) and site. Results: A total of 85,865 consecutive screening mammograms across the three clinical sites were identified. After controlling for age and race, adjusted odds ratios (aOR) of a mammogram being classified as dense at Site C compared to Site B before the DL model was implemented was 2.01 (95% CI 1.873, 2.157, p<0.001). This increased to 2.827 (95% CI 2.636, 3.032, p< 0.001) after DL implementation. The aOR of a mammogram being classified as dense at Site A after implementation compared to before implementation was 0.924 (95% CI 0.885, 0.964, p<0.001). Conclusions: Our findings suggest implementation of the DL model influences radiologist’s prospective density assessments in routine clinical practice by reducing the odds of a screening exam being categorized as dense. As a result, clinical use of our model could reduce downstream costs of supplemental screening tests and limit unnecessary high-risk clinic evaluations.[Table: see text]
Deep Learning for Mammographic Breast Density Assessment and Beyond
